Re: [exim] High Perf server

On Thu, Jun 30, 2005 at 03:35:35PM +0100, Tony Finch wrote:
> In order to do that you have to use a single process (threaded or
> event-driven) server which is WAY different from Exim :-)

Or a distributed algorithm to coordinate things, but yes, it is very
different from the current code.

> > Ok, so you think the queue is not the largest bottleneck? That may well
> > be true in your environment. In mine, it is.
>
> Are you sure it's the queue and not the hints DBs? Are you using a ram
> disk for them? Have you set no_message_logs?

I am using tmpfs for hints. I will try no_message_logs, but as I
said: Removing the fsync() calls already helps a lot, probably by
avoiding disk transfers for messages that are delivered and deleted
right after being received. I don't think msglog is being fsync()ed.
Although creating and removing it modifies the directory and the inode
and block allocation structures, it does so continuosly, combining all
modifications and writing them back to disk every now and then, needing
only few transfers in total and most likely by far less than one per
message.

Perhaps msglog should be another tmpfs in order to keep the logs,
but not having any disk I/O. I'll try that, too.

Michael